What is repwam?

RepWAM is the code repository for an academic paper about teaching robots to understand and predict how the world changes when actions happen. It comes from researchers at Fudan University and Ant Group's Robbyant lab, and pairs with a paper posted on arXiv and a Hugging Face listing. The approach has two parts. First, the team trains what they call a visual-action tokenizer, named RepViTok, which watches video and learns to compress both the visual scene and the actions that connect one moment to the next into compact tokens. It is trained with both plain pixel reconstruction and a semantic alignment step, meaning it is pushed to capture meaningful scene content rather than just raw pixel detail. Second, on top of that tokenizer, they train what they call a world action model, a system that predicts future visual states and the actions linking them, guided by written language instructions. The stated goal is to let a system that learns general patterns about how the visual world evolves transfer that understanding into controlling a real robot. The README reports results on a real Franka dual arm robot across three manipulation tasks, such as picking fruit into a plate, pushing open a drawer, and inserting a test tube into a rack, claiming the model outperforms other vision language action systems and prior world action models on these. It also reports scores on a benchmark called RoboTwin 2.0 across 50 tasks, and states that swapping in their tokenizer in place of an existing one improves average success rates. The README notes an open source plan where the paper itself has already been released, with inference code and full code and model weights listed as still to come as of this writing, so the repository may not yet contain runnable code for everyone. No license file or terms are mentioned in the README.